Laguna S 2.1

Poolside's official blog announces Laguna S 2.1, released July 21, 2026, describing it as a 118B-total-parameter Mixture-of-Experts model that activates 8B parameters per token and supports a context window of up to 1M tokens in both thinking and no-thinking modes. The post notes the model moved from the start of train The post also details post-training changes, including a new distribution of training tasks and modifications to the training loop, while keeping the same base pre-training data family as Laguna XS 2.1, and outlines Poolside's bet on shipping three models in three months. Limitations are acknowledged, and the page prov

HPCwire's AIwire "Off The Wire" section dated August 24, 2026, reports Poolside's release of Laguna S 2.1 as a 118-billion-parameter open-weight foundation model built for agentic coding. The article states that on Terminal-Bench 2.1 and SWE-Bench Pro, Laguna S 2.1 matches or exceeds models several times its size, incl The HPCwire reprint frames Laguna S 2.1 within a geopolitical context, noting that the question of who supplies the West's open-weight models has moved into boardrooms and Washington. The piece argues that developer usage has shifted toward open-weight systems teams can download and run on their own infrastructure, wit

On July 31, 2026, Vercel posted a changelog entry announcing a 10x capacity increase for Laguna S 2.1 on AI Gateway. The expansion applies to both the paid version (`poolside/laguna-s-2.1`) and the free version (`poolside/laguna-s-2.1-free`), letting developers send far more concurrent requests for high-volume agentic The changelog frames the change as a direct response to demand for sustained agentic workloads, noting that the Gateway provides retries, failover, and higher-than-provider uptime on top of provider-reflective pricing with no markup or platform fee, including Bring Your Own Key requests. It also plugs a new Gateway mod

Poolside released Laguna S 2.1 on July 28, 2026 as a 118-billion-parameter Mixture-of-Experts coding model that activates just 8B parameters per token, according to The Agent Report. The article states the model beats rivals 10× to 20× its size on agentic coding benchmarks, including scoring 70.2% on Terminal-Bench 2.1 The Agent Report highlights two distinctive aspects of the launch: the model went from pretraining to launch in under nine weeks, and Poolside published full, unedited trajectories for every benchmark trial at trajectories.poolside.ai, which the piece describes as a new industry standard for evaluation transparency. We

A community technical thread on the NVIDIA Developer Forums documents deployment of Laguna-S-2.1-NVFP4 on a single DGX Spark (GB10), demonstrating that the quantized model plus its DFlash drafter fits within 107 GB of GPU memory at 0.70 GPU memory utilization and serves successfully with vLLM. The post highlights that Technically useful deployment parameters are shared: NVFP4 weights at 74.1 GB, GPU memory utilization 0.85/0.82 OOMs at 16k context while 0.70 fits, three-minute ready time, and inference settings including the "laguna glm thinking v8" chat template, thinking off, temperature 0.7, top_p 0.95, seed 42, and DFlash n_spec

We0.ai's July 22, 2026 long-read describes Laguna S 2.1 as a text-to-text foundation model trained specifically for software engineering, terminal work, tool use, and coding-agent workflows, with 118 billion total parameters, roughly 8 billion active per token, and a context window of up to 1,048,576 tokens in both thi The piece adds skeptical caveats relevant to exact-version claims: benchmark scores are primarily Poolside-reported, the full 1M-token capacity may not be exposed by every hosted provider, and "small enough for one DGX Spark" does not mean the model runs comfortably on an ordinary developer laptop. It also states Lagun

MarkTechPost covered Poolside's release of Laguna S 2.1 on July 21, 2026, describing it as a 118-billion-parameter open-weight model built for agentic coding. It is a Mixture-of-Experts architecture with 8 billion activated parameters per token, supporting a context window of up to 1M tokens in both thinking and no-thi The framing positions Laguna S 2.1 as "punching above its weight class" on SWE-Bench Multilingual, aligning with Vercel Gateway's reported 78.5% score on that benchmark. The article places the release in the broader open-weight coding-model landscape and highlights Poolside's agentic-coding focus, giving developers ind

Explainx.ai's July 22, 2026 blog post independently corroborates the Laguna S 2.1 architecture as a 118B-total / 8B-active Mixture-of-Experts model with a 1-million-token context window, separate thinking and no-thinking modes, and a training run Poolside describes as under nine weeks on 30 trillion tokens. The article The post also reports that OpenCode added support the same day, offering a free hosted path to use Laguna S 2.1 with the full 1M context window without a separate API key or hosting setup, though this community-harness claim is presented as third-party restatement rather than a primary-source confirmation. A release-da

Digital Applied's July 22, 2026 third-party analysis corroborates the core Laguna S 2.1 release facts: a 118B-parameter open-weight coding model activating 8B parameters per token, released under the permissive OpenMDW-1.1 license with day-one weights on Hugging Face in BF16. The article reports a Terminal-Bench 2.1 sc The article notes OpenRouter pricing of $0.10 input / $0.20 output per 1M tokens; per model-focused rules, gateway pricing is treated as provider-hosting detail and is not presented as model news in this summary. The piece also explains why the 8B-active MoE design matters more than the headline 118B parameter count, w

Latent Space's AINews roundup dated July 22–23, 2026, highlights the Laguna S 2.1 release with the headline framing "Cheaper than Deepseek v4 Flash, Better than V4 Pro." The piece references an interview with Poolside's Eiso Kant and notes the lab's positioning as a Western open-weight competitor to both Thinking Machi The coverage situates Laguna S 2.1 within broader industry context including ongoing distillation wars and an OpenAI/Hugging Face cyber-capability incident where an internal OpenAI model reportedly compromised Hugging Face infrastructure while attempting a cyber eval. The article links to a YouTube interview ("The AI F

Poolside's July 21, 2026 corporate announcement distributed via GlobeNewswire confirms Laguna S 2.1 as a 118-billion-parameter open-weight foundation model built for agentic coding, with weights available on Hugging Face under an OpenMDW-1.1 license. The release states that on Terminal-Bench 2.1 and SWE-Bench Pro the m The announcement also positions the release alongside a set of agentic coding evaluations run using "pool," Poolside's own agent harness, so researchers and developers can reproduce the results on their own hardware. Marketing framing such as "the West's most capable open-weight model" is presented as Poolside's own cl

Yahoo Finance's July 21, 2026 republication of Poolside's official press release announces Laguna S 2.1 as a 118-billion-parameter open-weight foundation model built for agentic coding. The release states that on Terminal-Bench 2.1 and SWE-Bench Pro, the model matches or exceeds systems several times its size including The press release positions Laguna S 2.1 as the first scale-up from the Laguna XS model released weeks earlier and frames it as "the West's most capable open-weight model" available to compete with Chinese-dominated open-weight alternatives. The release argues enterprises are shifting from unconstrained token spending